Spaces:
Sleeping
Sleeping
小形克宏 commited on
Commit ·
f613e51
1
Parent(s): 92b319a
Add dataset format distribution analyzer
Browse files- dataset_analyzer.py +211 -0
dataset_analyzer.py
ADDED
|
@@ -0,0 +1,211 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""
|
| 2 |
+
Dataset Format Analyzer
|
| 3 |
+
SFTデータセットのフォーマット分布を分析するスクリプト
|
| 4 |
+
|
| 5 |
+
指定されたHuggingFaceデータセットをダウンロードし、
|
| 6 |
+
各サンプルのターゲット出力がどのフォーマット(JSON/YAML/TOML/XML/CSV)
|
| 7 |
+
であるかを判定・集計します。
|
| 8 |
+
"""
|
| 9 |
+
|
| 10 |
+
import json
|
| 11 |
+
import re
|
| 12 |
+
import sys
|
| 13 |
+
from collections import Counter, defaultdict
|
| 14 |
+
|
| 15 |
+
def detect_format(text):
|
| 16 |
+
"""テキストのフォーマットを推定する"""
|
| 17 |
+
text = text.strip()
|
| 18 |
+
|
| 19 |
+
# マークダウンブロック除去
|
| 20 |
+
cleaned = re.sub(r"```\w*\n?", "", text).strip()
|
| 21 |
+
if not cleaned:
|
| 22 |
+
return "EMPTY"
|
| 23 |
+
|
| 24 |
+
# JSON: { or [ で始まる
|
| 25 |
+
if cleaned.startswith("{") or cleaned.startswith("["):
|
| 26 |
+
try:
|
| 27 |
+
json.loads(cleaned)
|
| 28 |
+
return "JSON"
|
| 29 |
+
except:
|
| 30 |
+
return "JSON" # JSONっぽいが壊れている
|
| 31 |
+
|
| 32 |
+
# XML: < で始まる(<?xml or <tag)
|
| 33 |
+
if cleaned.startswith("<"):
|
| 34 |
+
return "XML"
|
| 35 |
+
|
| 36 |
+
# CSV: カンマ区切りの複数行
|
| 37 |
+
lines = cleaned.split("\n")
|
| 38 |
+
if len(lines) >= 2:
|
| 39 |
+
comma_counts = [line.count(",") for line in lines[:5] if line.strip()]
|
| 40 |
+
if comma_counts and all(c == comma_counts[0] and c > 0 for c in comma_counts):
|
| 41 |
+
return "CSV"
|
| 42 |
+
|
| 43 |
+
# TOML: [section] パターンまたは key = value パターン
|
| 44 |
+
if re.match(r"^\[[\w\.\-]+\]", cleaned) or re.match(r'^[\w\.\-]+\s*=\s*', cleaned):
|
| 45 |
+
return "TOML"
|
| 46 |
+
|
| 47 |
+
# YAML: key: value パターン(インデント構造)
|
| 48 |
+
if re.match(r'^[\w\-]+:\s', cleaned) or cleaned.startswith("---") or cleaned.startswith("- "):
|
| 49 |
+
return "YAML"
|
| 50 |
+
|
| 51 |
+
return "OTHER"
|
| 52 |
+
|
| 53 |
+
|
| 54 |
+
def detect_format_from_prompt(prompt_text):
|
| 55 |
+
"""プロンプト(query)からターゲットフォーマットを推定"""
|
| 56 |
+
prompt_lower = prompt_text.lower()
|
| 57 |
+
|
| 58 |
+
# 明示的な指示を検索
|
| 59 |
+
patterns = {
|
| 60 |
+
"JSON": [r"output\s+json", r"to\s+json", r"in\s+json", r"json\s+code", r"json\s+format"],
|
| 61 |
+
"YAML": [r"output\s+yaml", r"to\s+yaml", r"in\s+yaml", r"yaml\s+code", r"yaml\s+format"],
|
| 62 |
+
"TOML": [r"output\s+toml", r"to\s+toml", r"in\s+toml", r"toml\s+code", r"toml\s+format"],
|
| 63 |
+
"XML": [r"output\s+xml", r"to\s+xml", r"in\s+xml", r"xml\s+code", r"xml\s+format"],
|
| 64 |
+
"CSV": [r"output\s+csv", r"to\s+csv", r"in\s+csv", r"csv\s+code", r"csv\s+format"],
|
| 65 |
+
}
|
| 66 |
+
|
| 67 |
+
for fmt, pats in patterns.items():
|
| 68 |
+
for pat in pats:
|
| 69 |
+
if re.search(pat, prompt_lower):
|
| 70 |
+
return fmt
|
| 71 |
+
|
| 72 |
+
# タスク名パターン (e.g., "Text to JSON", "CSV to YAML")
|
| 73 |
+
task_pattern = r"(text|json|yaml|toml|xml|csv)\s+to\s+(json|yaml|toml|xml|csv)"
|
| 74 |
+
match = re.search(task_pattern, prompt_lower)
|
| 75 |
+
if match:
|
| 76 |
+
return match.group(2).upper()
|
| 77 |
+
|
| 78 |
+
return None
|
| 79 |
+
|
| 80 |
+
|
| 81 |
+
def analyze_dataset(dataset_id):
|
| 82 |
+
"""HuggingFaceデータセットを分析"""
|
| 83 |
+
from datasets import load_dataset
|
| 84 |
+
|
| 85 |
+
print(f"📥 データセットをダウンロード中: {dataset_id}")
|
| 86 |
+
ds = load_dataset(dataset_id, split="train")
|
| 87 |
+
print(f"✅ ダウンロード完了: {len(ds)} 件\n")
|
| 88 |
+
|
| 89 |
+
# messages構造を解析
|
| 90 |
+
format_from_output = Counter()
|
| 91 |
+
format_from_prompt = Counter()
|
| 92 |
+
task_types = Counter()
|
| 93 |
+
cot_count = 0
|
| 94 |
+
samples_by_format = defaultdict(list)
|
| 95 |
+
|
| 96 |
+
for i, row in enumerate(ds):
|
| 97 |
+
messages = row.get("messages", [])
|
| 98 |
+
|
| 99 |
+
# messagesからuser/assistantを抽出
|
| 100 |
+
user_msg = ""
|
| 101 |
+
assistant_msg = ""
|
| 102 |
+
has_cot = False
|
| 103 |
+
|
| 104 |
+
for msg in messages:
|
| 105 |
+
role = msg.get("role", "")
|
| 106 |
+
content = msg.get("content", "")
|
| 107 |
+
if role == "user":
|
| 108 |
+
user_msg = content
|
| 109 |
+
elif role == "assistant":
|
| 110 |
+
assistant_msg = content
|
| 111 |
+
if "<think>" in content or "</think>" in content:
|
| 112 |
+
has_cot = True
|
| 113 |
+
|
| 114 |
+
if has_cot:
|
| 115 |
+
cot_count += 1
|
| 116 |
+
|
| 117 |
+
# CoT部分を除去してアシスタントの最終出力を取得
|
| 118 |
+
final_output = assistant_msg
|
| 119 |
+
think_match = re.search(r"</think>\s*(.*)", assistant_msg, re.DOTALL)
|
| 120 |
+
if think_match:
|
| 121 |
+
final_output = think_match.group(1).strip()
|
| 122 |
+
|
| 123 |
+
# 出力フォーマットを判定(2つの方法)
|
| 124 |
+
fmt_output = detect_format(final_output)
|
| 125 |
+
fmt_prompt = detect_format_from_prompt(user_msg)
|
| 126 |
+
|
| 127 |
+
format_from_output[fmt_output] += 1
|
| 128 |
+
if fmt_prompt:
|
| 129 |
+
format_from_prompt[fmt_prompt] += 1
|
| 130 |
+
else:
|
| 131 |
+
format_from_prompt["UNKNOWN"] += 1
|
| 132 |
+
|
| 133 |
+
# タスクタイプ推定
|
| 134 |
+
task_match = re.search(r"(text|json|yaml|toml|xml|csv)\s+to\s+(json|yaml|toml|xml|csv)", user_msg.lower())
|
| 135 |
+
if task_match:
|
| 136 |
+
task_type = f"{task_match.group(1).upper()} to {task_match.group(2).upper()}"
|
| 137 |
+
elif "please output" in user_msg.lower():
|
| 138 |
+
task_type = f"Text to {fmt_prompt or fmt_output}"
|
| 139 |
+
else:
|
| 140 |
+
task_type = "OTHER"
|
| 141 |
+
task_types[task_type] += 1
|
| 142 |
+
|
| 143 |
+
# サンプル保存(各フォーマット最大2件)
|
| 144 |
+
fmt_key = fmt_prompt or fmt_output
|
| 145 |
+
if len(samples_by_format[fmt_key]) < 2:
|
| 146 |
+
samples_by_format[fmt_key].append({
|
| 147 |
+
"index": i,
|
| 148 |
+
"prompt_preview": user_msg[:150],
|
| 149 |
+
"output_preview": final_output[:150],
|
| 150 |
+
})
|
| 151 |
+
|
| 152 |
+
# --- 結果出力 ---
|
| 153 |
+
total = len(ds)
|
| 154 |
+
|
| 155 |
+
print("=" * 70)
|
| 156 |
+
print(f"📊 データセット分析結果: {dataset_id}")
|
| 157 |
+
print(f" 総サンプル数: {total}")
|
| 158 |
+
print(f" CoTあり: {cot_count} ({cot_count/total*100:.1f}%)")
|
| 159 |
+
print("=" * 70)
|
| 160 |
+
|
| 161 |
+
print(f"\n📋 ターゲットフォーマット分布(プロンプトから判定):")
|
| 162 |
+
print(f"{'Format':<12} {'Count':>6} {'Percent':>8}")
|
| 163 |
+
print("-" * 30)
|
| 164 |
+
for fmt in ["JSON", "YAML", "TOML", "XML", "CSV", "UNKNOWN"]:
|
| 165 |
+
count = format_from_prompt.get(fmt, 0)
|
| 166 |
+
pct = f"{count/total*100:.1f}%"
|
| 167 |
+
bar = "█" * int(count/total*50)
|
| 168 |
+
print(f"{fmt:<12} {count:>6} {pct:>8} {bar}")
|
| 169 |
+
|
| 170 |
+
print(f"\n📋 出力フォーマット分布(出力内容から判定):")
|
| 171 |
+
print(f"{'Format':<12} {'Count':>6} {'Percent':>8}")
|
| 172 |
+
print("-" * 30)
|
| 173 |
+
for fmt, count in format_from_output.most_common():
|
| 174 |
+
pct = f"{count/total*100:.1f}%"
|
| 175 |
+
bar = "█" * int(count/total*50)
|
| 176 |
+
print(f"{fmt:<12} {count:>6} {pct:>8} {bar}")
|
| 177 |
+
|
| 178 |
+
print(f"\n📋 タスクタイプ分布:")
|
| 179 |
+
print(f"{'Task Type':<25} {'Count':>6} {'Percent':>8}")
|
| 180 |
+
print("-" * 45)
|
| 181 |
+
for task, count in task_types.most_common(20):
|
| 182 |
+
pct = f"{count/total*100:.1f}%"
|
| 183 |
+
print(f"{task:<25} {count:>6} {pct:>8}")
|
| 184 |
+
|
| 185 |
+
# public_150との比較
|
| 186 |
+
print(f"\n📋 public_150.json との比較(参考):")
|
| 187 |
+
print(f"{'Format':<8} {'public_150':>12} {'dataset':>12} {'充足度':>10}")
|
| 188 |
+
print("-" * 45)
|
| 189 |
+
public_counts = {"JSON": 50, "YAML": 35, "TOML": 25, "XML": 20, "CSV": 20}
|
| 190 |
+
for fmt in ["JSON", "YAML", "TOML", "XML", "CSV"]:
|
| 191 |
+
pub = public_counts[fmt]
|
| 192 |
+
ds_count = format_from_prompt.get(fmt, 0)
|
| 193 |
+
ratio = f"{ds_count/pub:.1f}x" if pub > 0 else "N/A"
|
| 194 |
+
print(f"{fmt:<8} {pub:>12} {ds_count:>12} {ratio:>10}")
|
| 195 |
+
|
| 196 |
+
print(f"\n📋 各フォーマットのサンプル:")
|
| 197 |
+
for fmt in ["JSON", "YAML", "TOML", "XML", "CSV"]:
|
| 198 |
+
samples = samples_by_format.get(fmt, [])
|
| 199 |
+
print(f"\n--- {fmt} サンプル ({len(samples)}件) ---")
|
| 200 |
+
for s in samples:
|
| 201 |
+
print(f" [#{s['index']}] prompt: {s['prompt_preview'][:100]}")
|
| 202 |
+
print(f" output: {s['output_preview'][:100]}")
|
| 203 |
+
|
| 204 |
+
|
| 205 |
+
if __name__ == "__main__":
|
| 206 |
+
if len(sys.argv) > 1:
|
| 207 |
+
dataset_id = sys.argv[1]
|
| 208 |
+
else:
|
| 209 |
+
dataset_id = "u-10bei/structured_data_with_cot_dataset_512_v4"
|
| 210 |
+
|
| 211 |
+
analyze_dataset(dataset_id)
|